Glaucoma is related to the progressive degeneration of optic nerve fibers and ... volutional neural network architecture merged with cup-to-disc morphological .... The learning rate (lr) parameter was explored for each model using a grid search.
Combining morphometric features and convolutional networks fusion for glaucoma diagnosis a
a
a
Oscar Perdomo , John Arevalo , Fabio A. González a
MindLab Research Group, Universidad Nacional de Colombia, Bogotá
Glaucoma is an eye condition that leads to loss of vision and blindness. Ophthalmoscopy exam evaluates the shape, color and proportion between the optic disc and physiologic cup, but the lack of agreement among experts is still the main diagnosis problem. The application of deep convolutional neural networks combined with automatic extraction of features such as: the cup-to-disc distance in the four quadrants, the perimeter, area, eccentricity, the major radio, the minor radio in optic disc and cup, in addition to all the ratios among the previous parameters may help with a better automatic grading of glaucoma. This paper presents a strategy to merge morphological features and deep convolutional neural networks as a novel methodology to support the glaucoma diagnosis in eye fundus images. Keywords: Glaucoma diagnosis, feature-fusion, morphometry, eye fundus
image, feature-learning.
1
INTRODUCTION
The diagnosis of eye condition is done with dierent kind of clinical exams. Exams may be non-invasive such as: slit-lamp exam, visual acuity, fundus eye image, ultrasound, optical coherence tomography (OCT); or invasive exams as uoresce in angiography [1]. the non-invasive clinical exams are easier to make, have no contraindications and do not aect the eye's natural response to external factors in comparison to the invasive exams. Therefore, eye fundus image is a high patient compliance, quick and simple technique, with the main advantages that in most cases dilation is not required, images can be easily saved to be analyzed at a later time, and progression of diseases can be monitored over time. Glaucoma is related to the progressive degeneration of optic nerve bers and structural changes of the optic nerve head [1]. Although glaucoma cannot be cured, its progression can be slowed down by treatment. Therefore, timely diagnosis of this disease is vital to avoid blindness [2]. Glaucoma diagnosis detection
1
is based on manual assessment of the Optic Disc (OD) through ophthalmoscopy and posterior eye fundus image analysis, looking morphological parameters for the central bright zone called the physiologic cup and a peripheral region called the neuroretinal rim [1, 2]. The automatic cup-to-disc ratio (CDR) in eye fundus images has been used as the main physiological characteristics in the diagnosis of glaucoma [3]. Some researchers have been focusing on global and regional features such as texture, grayscale and wavelet energy of the Optic Nerve Head (ONH) to classify normal and glaucoma images [4]. Other study focused in texture property of the total image using Haralick features plus neural networks [5].
Deep learning archi-
tecture has been explored for other eye conditions diagnosis [6] and automated glaucoma diagnosis.
Sevastopolsky used a modied U-Net CNN on publicly
available eye fundus images DRIONS-DB, RIM-ONE v.3, DRISHTI-GS to do optic disc and cup segmentation [7]. Chen et al. built a six CNN layers to get best performance of glaucoma diagnosis [8]. Orlando et al. studied pre-trained OverFeat and VGG-S CNN from non-medical data applied to eye fundus images in order to detect glaucoma [9]. Despite the good results obtained, these studies were not assessed the ability of the models to classify the eye fundus images as healthy, suspicious and glaucoma on a common image database, and thus the performance numbers are not directly comparable, making it dicult to assess the true performance of these methods. The aim of the present work is to explore the strategy of using a deep convolutional neural network architecture merged with cup-to-disc morphological features to improve the classication of healthy, suspicious and glaucoma images. The remainder of this paper is organized as follows: First, in Section 2, we give an overview of the proposed method including automatic extraction of morphological features and the deep convolutional neural network architecture. Then, in Section 3 we describe the experimental setup using as reference the baseline in order to build the dataset, and the dierent experiments along with the performance achieved. In Section 4, the results are presented and discussed. Finally, Section 5 presents the conclusions and future work.
2
Methods
The proposed method, is depicted in Figure 1. In the rst stage, 19 morphological features are extracted using disc and cup segmentation. The second stage learns a set of features using a deep convolutional neural network (DCNN). The nal stage combines both, morphological and convolutional features merging them to feed the loss function. Kappa loss function is preferred over the traditional softmax function since, there is a relation between grades of diagnosis [10]. The model is trained jointly by applying stochastic gradient descent.
2
Figure 1:
Block diagram used to classify glaucoma condition in eye fundus
images
2.1
Automatic extraction of morphological features of eye fundus images
It has been shown that the eye morphometry in fundus images helps to glaucoma diagnosis [3]. This work proposed a set of 19 morphometric features based on the optic disc and physiologic cup segmentations. The rst step is to extract the four quadrants from the image as shown in Figure 2.
Figure 2: [left] Four quadrants and cup-to-disc ratios in an eye fundus image of a subject with glaucoma. [center] Optic disc segmentation performed by expert, and [right] Physiologic cup segmentation performed by expert. Secondly, the cup-to-disc distance for each quadrant was calculated. Also, perimeter, area, eccentricity, the major radio and the minor radio were calculated for both, optic disc and cup. Additionally, 5 ratios were included in order
3
Superior distance optic disc
Perimeter optic disc
Inferior distance optic disc
Eccentricity optic disc
Eccentricity physiologic cup Cup-to-disc area ratio
Temporal distance optic disc
Area physiologic cup
Cup-to-disc major axis ratio
Nasal distance optic disc
Major axis physiologic cup
Cup-to-disc minor axis ratio
Area disc
Minor axis physiologic cup
Cup-to-disc perimeter ratio
Major axis optic disc
Perimeter physiologic cup
Cup-to-disc eccentricity ratio
Minor axis optic disc
Table 1: List of morphometric measures extracted from disc and cup segmentation.
to capture disproportions between optic disc and cup. Table 1 summarizes the 19 morphometric features proposed in this work.
2.2
Deep Convolutional Neural Network
Deep Convolutional Neural Network (DCNN) is a model designed with a big number of layers to learn a representation of data containing spatial relations. This is the case of eye fundus images, where spatial patterns are determinant to diagnose dierent eye diseases, e.g. glaucoma, making the DCNN a suitable approach for image classication. DCNN learns a set of features using a minimal preprocessing while, with the properly supervised training, may respond to distortion, variability and invariant patterns. The DCNN is composed of 5 convolutional layers with kernel size of layers with pool size of padding of
1 × 1,
2×2
3×3
and stride of
and strides of
2 × 2,
1 × 1,
5 max-pooling
4 zero-padding layers with
and 2 fully-connected layers with 512 and x-class number un-
tis. A convolutional layer is composed of a set of learnable lters that convolved with the input generating an activation map for each lter. The convolutional layer output is the input of a max pooling layer that is a non-linear size reducer that is applied to the activation choosing the maximum value of a set of contiguous pixels. A zero-padding layer adds a set of pixels of value 0 to increase the image size but without aect the image information, these layers were applied in order to ensure an even dimension at max-pooling layer's output.
Finally,
the fully-connected layer connects all the neurons in the previous layer to the next layer. The DCCN architecture used in this work is described in Table 2.
3
Experimental setup
3.1
RIM-ONE-r3 dataset
The RIM-ONE-r3[11] database with eye color fundus images was used in this study.
The database contains
images from healthy subjects, sis and
39
159 images with size of 1072 × 1424 pixels, 85 35 images with a suspected glaucoma diagno-
images with glaucoma diagnosis. The images were labelled by two
ophthalmologist experts from the Deparment of Ophthalmology at the Hospital
4
N
Name
Channels
Width
Height
Filter size
Stride
0
Input
3
224
224
-
-
1
Padding1
3
226
226
-
-
2
Conv1
32
224
224
3x3
1x1
3
Max Pool1
32
112
112
2x2
2x2
4
Padding2
32
114
114
-
-
5
Conv2
64
112
112
3x3
1x1
6
Max Pool2
64
56
56
2x2
2x2
7
Conv3
64
54
54
3x3
1x1
8
Padding3
64
56
56
-
-
9
Max Pool3
64
28
28
2x2
2x2
10
Padding4
64
30
30
-
-
11
Conv4
64
28
28
3x3
1x1
12
Max Pool4
64
14
14
2x2
2x2
13
Conv5
64
12
12
3x3
1x1
14
Max Pool5
64
6
6
2x2
2x2
15
Fully Conn1
512
-
-
-
-
16
Fully Conn2
num_classes
-
-
-
-
17
Kappa_loss
num_classes
-
-
-
-
Table 2: Arquitecture of the DCNN with values used in each layer.
Universitario de Canarias in Spain [10]. The proposed method was evaluated in 2 setups: an unbalanced 3-class classication setup (healthy, suspicious and glaucoma) and a binary classication setup (healthy vs. suspicious+glaucoma), this was created to balance the classes and to assess the detection ability of models to discriminate between healthy class vs non-healthy class. The dataset was randomly split in a patient basis training (60%), validation (10%) and test (30%) subsets with stratied sampling.
3.2
Evaluation
Two congurations of the proposed model were evaluated a conguration using only the convolutional network (DCNN) and a conguration that also includes the morfometric features (DCNN + MFs). The models were trained using stochastic gradient descent on both the 2-classes and the 3-classes problems.. The learning rate (lr) parameter was explored for each model using a grid search strategy, the best performing values found in validation are listed in Table 3, using
200 as the number of epochs to train the model.
The 18-layers DCNN was
1 using
chosen as baseline. The proposed approach was implemented with Keras
GeForce GTX TITAN X from NVIDIA. The Kappa coecient was implemented as a cost function, and loss, precision, recall, f-score and Kappa measures were reported for both training and test sets.
1 http://keras.io
5
SVM and RF were evaluated as baseline methods and trained using morphometric features normalized with (mean
= 0)
and (variance
= 1). C
parameter
for the linear SVM was explored.
4
Results
Experimental results are reported in Table 3.
The best performance of the
proposed model was obtained with a learning rate of learning rate of
Method
0.0001
0.01
for two classes and a
for three classes both with a batch size of
Num_Classes 2 Classes
DCNN 3 Classes 2 Classes DCNN + MFs 3 Classes
32.
lr Precision Recall f-score Kappa 0.1 0.0 0.0 0.0 0.0 0.01 0.59 0.77 0.67 0.23 0.001 0.55 0.23 0.32 0.0 0.0001 0.50 0.05 0.08 0.0 0.1 0.17 0.33 0.23 0.0 0.01 0.17 0.33 0.23 0.0 0.001 0.17 0.33 0.23 0.0 0.0001 0.17 0.33 0.23 0.0 0.1 0.74 0.77 0.76 0.52 0.01
0.90
0.86
0.88
0.78
0.0001
0.46
0.56
0.50
0.42
0.001 0.0001 0.1 0.01 0.001
0.81 0.80 0.47 0.42 0.49
0.77 0.73 0.47 0.40 0.54
0.79 0.76 0.47 0.41 0.51
0.61 0.54 0.27 0.08 0.39
Table 3: Performance of the two models with dierent learning rates in validation dataset. [In boldface] the best performance achieved at 2 setups. We evaluated the proposed model with the best parameters applied to the test dataset. Table 3 presents the precision, recall, macro averaged f-score and Kappa coecient results of the proposed methods compared with baseline methods. The proposed method clearly outperforms SVM and RF in Kappa coecient that measures inter-rater agreement among the binary classication (balanced setup) and 3-class problems (unbalanced setup) [12].
This showed the
proposed method is able to capture the visual features and morphological features that characterize glaucoma and combine them to improve the glaucoma diagnosis.
6
Method
Num_Classes
Precision
Recall
f-score
Kappa
0.74
0.77
0.75
0.52
0.88
0.68
0.76
0.61
0.90
0.86
0.88
0.78
0.33
SVM RF
2 Classes
Proposed method SVM RF
3 Classes
Proposed method
0.63
0.56
0.55
0.64
0.57
0.58
0.35
0.46
0.56
0.50
0.42
Table 4: Performance measures in the baseline models and the proposed method in test dataset. [In boldface] the best performance achieved at 2 setups.
5
Discussion and conclusion
Experimental results showed that the DCNN model combined with morphological features is highly correlated with the three classes ground truth to classify glaucoma condition according to Kappa coecient.
As shown in table 4 the
precision and f-score are lower than our propossed method and this is due to the unbalancing on the classes, as healthy class doubled in number of samples for the other two classes the SVM and RF classiers went biased towards classify this class. For this reasson we also measured the performance in terms of Kappa coecient. This coecient is widely used in medicine to compare classication performances regardless balanced or unbalanced setups [12]. According to [12] the Kappa presented at the two classication problems presented good (0.60 to 0.80) and moderate (0.40 to 0.60) agreements respectly. The main advantage of the proposed method is that it uses as input the raw image and the optic disc and physiologic cup segmentations to calculate morphologic features merging those two sources of information in a vector to improve the glaucoma classication. The automatic extraction of morphological features related to optic disc and physiologic cup may improve the disease diagnosis, but the combination of DCNN with this kind of approach showed good preliminary results in glaucoma detection. Its application to other datasets is the subject of our future work.
Acknowlegement Oscar Perdomo and John Arévalo thank COLCIENCIAS for funding this research with a doctoral grant.
References [1] Zhang, Z., Ruchir, S.H.L., Xiangyu C.L.D., Damon W. K.W., Chee K.K., Tien Y.W., and Jiang L., A survey on computer aided diagnosis for ocular diseases, BMC medical informatics and decision making, 14(1), 1-29 (2014).
7
[2] Bock, R, Meier, J., Nyúl, L., and Michelson, G., Glaucoma risk index: automated glaucoma detection from color fundus images, Med Image Anal., 14(3), 471-481 (2010). [3] Ranjith, N., Saravanan, C., and Bibin, M.R., Glaucoma Diagnosis by Optic Cup to Disc Ratio Estimation, International Journal of Inventive Engineering and Sciences (IJIES), 3(5), 1-5 (2015). [4] Mookiah, M.R.K., Acharya, U.R., Lim, C.M., Petznick A, and Suri J.S., Data mining technique for automated diagnosis of glaucoma using higher order spectra and wavelet energy features, Knowledge-Based Syst., 33, 73-82 (2012). [5] Samanta, S., Ahmed, S.S., Salem, M.A.M.M., Nath, S.S., Dey, N., and Chowdhury, S.S., "Haralick Features Based Automated Glaucoma Classication Using Back Propagation Neural Network," In FICTA (1), 351-358 (2014). [6] Perdomo, O., Otalora, S., Rodríguez, F., Arevalo, J., and González, F. A., A Novel Machine Learning Model Based on Exudate Localization to Detect Diabetic Macular Edema, OMIA 2016, Held in Conjunction with MICCAI 2016, 137-144 (2016). [7] Sevastopolsky, A., "Optic Disc and Cup Segmentation Methods for Glaucoma Detection with Modication of U-Net Convolutional Neural Network," arXiv preprint arXiv:1704.00979 (2017). [8] Chen, X., Xu, Y., Wong, D.W.K., Wong, T.Y., and Liu, J., "Glaucoma detection based on deep convolutional neural network," In Engineering in Medicine and Biology Society (EMBC), 37th Annual International Conference of the IEEE, 715-718 (2015). [9] Orlando, J.I., Prokofyeva, E., del Fresnob, M., and Blaschko, M., "Convolutional neural network transfer for automated glaucoma identication," In 12th International Symposium on Medical Information Processing and Analysis, International Society for Optics and Photonics, 101600U-101600U (2017). [10] Perdomo, O., Arevalo, J., and González, F. A., Convolutional network to detect exudates in eye fundus images of diabetic subjects, In 12th International Symposium on Medical Information Processing and Analysis, International Society for Optics and Photonics, 101600T-101600T (2017). [11] Fumero, F., Alayón, S., Sanchez, J.L., Sigut, J., and Gonzalez-Hernandez, M., "RIM-ONE: An open retinal image database for optic nerve evaluation," In Computer-Based Medical Systems (CBMS), 24th International Symposium, 1-6 (2011). [12] Altman D. G., [Practical Statistics for Medical Research], CRC Press, London, 404-408 (1990).
8